Tasks & Data — RoCo-Spring

Tracks

RoCo-Spring tasks

RoCo-Spring includes three quantitative dense correspondence tracks and one exploration track. The quantitative tracks evaluate methods on optical flow, stereo matching, and scene flow using paired clean and corrupted benchmark data. The exploration track rewards careful analysis of robustness behavior, failure modes, and evaluation methodology.

Quantitative tracks

Optical Flow

Estimate a dense 2D displacement field between consecutive frames, assigning each visible pixel a motion vector.

Stereo Matching

Estimate a dense disparity map from a rectified stereo image pair, assigning each pixel the horizontal offset to its corresponding point.

Scene Flow

Estimate dense 3D motion from stereo image sequences, combining per-pixel geometry and temporal motion into a single correspondence task.

Qualitative track

Exploration track

The exploration track is for rigorous studies that help explain robustness in dense correspondence. Submissions may analyze why current methods fail under corruptions, which design choices improve robustness, how different corruptions affect different model families, or whether robustness gains transfer across optical flow, stereo matching, and scene flow.

Strong exploration submissions should pose a clear scientific question, use a careful experimental protocol, provide reproducible evidence, and offer insights that are useful for future robust dense correspondence research.

Recommended exploration resource

FlowBench

Paper & benchmark

FlowBench is a useful starting point for exploration-track studies of optical-flow reliability and generalization. Its benchmark provides a broad, reproducible view of model behavior across clean data, adversarial perturbations, and common corruptions, making it valuable for selecting methods, forming hypotheses, and designing robust comparisons.

57 model checkpoints 3 benchmark datasets 9 adversarial attacks 23 common corruptions

Participants may build on FlowBench's study design or use its benchmark to investigate failure modes, reliability trade-offs, and whether conclusions generalize across model families and evaluation conditions.

If you use the FlowBench study, analyses, or benchmark, please cite:

@article{
agnihotri2025flowbench,
title={FlowBench: Benchmarking Optical Flow Estimation Methods for Reliability and Generalization},
author={Shashank Agnihotri and Julian Yuya Caspary and Luca Schwarz and Xinyan Gao and Jenny Schmalfuss and Andres Bruhn and Margret Keuper},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=Kh4bj6YDNm},
note={}
}

Datasets

Spring and RobustSpring

We use two datasets for RoCo-Spring: Spring provides clean high-resolution stereo video and dense ground truth for all three correspondence tasks; RobustSpring applies realistic corruptions to the same data for robustness evaluation. You can find links to the papers, benchmark, and datasets below.

RobustSpring applies corruptions to the same scenes, cameras, and labels as Spring. This paired clean / corrupted design measures robustness without changing scene content, camera setup, annotation protocol, or benchmark procedure. Performance drops can therefore be attributed to corruption sensitivity rather than unrelated dataset differences.

Spring

Spring provides high-resolution stereo video with left and right images at 1920x1080 px, together with dense scene-flow ground truth at 4x super-resolution. It includes bidirectional disparities for stereo and depth, disparity changes for scene flow, and forward/backward optical flow for both left and right views. This makes Spring a unified high-detail benchmark for optical flow, stereo matching, and scene flow.

Paper Benchmark Dataset

If you use the Spring dataset, please cite:

@inproceedings{mehl2023spring,
    title={Spring: A High-Resolution High-Detail Dataset and Benchmark for Scene Flow, Optical Flow and Stereo},
    author={Mehl, Lukas and Schmalfuss, Jenny and Jahedi, Azin and Nalivayko, Yaroslava and Bruhn, Andr{\'e}s},
    booktitle={Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    pages={4981--4991},
    year={2023}
}

RobustSpring

RobustSpring extends Spring into a robustness benchmark with 20 realistic image corruptions, including blur, color changes, noise, quality degradations, and weather effects. The corruptions are applied to stereo video data and are integrated consistently over time, across stereo views, and with depth where applicable, enabling controlled robustness evaluation for optical flow, stereo, and scene flow.